External return to education in Europe

(1)

Edukacja 2014, 6(131) An interdisciplinary approach ISSN 0239-6858 pp. 117–134

External return to education in Europe

Paweł Strawiński

Faculty of Economic Sciences, University of Warsaw*

In the literature, social return to education is defined as the sum of human capital return and external return.

The novelty of this study is that it provides an international comparison of external return to education. Many authors claim that the social return rate exceeds the pure technical rate of return by a considerable margin.

However, measurement of social return is challenged methodologically and by data problems. The approach employed in this study is based on comparative advantage theory which allows control for potential endogeneity and self-selection into different streams of education. External return was found to be positive in all European countries although magnitudes varied. The external return was greater in smaller economies where there was a smaller proportion of highly educated people.

Keywords: return to education, private returns, external returns, social returns.

A

lthough numerous studies have proved that investment in education is profita- ble at an individual level, not much is known about the profitability of tertiary education at an economic level. This study investigates human capital externalities in several Euro- pean countries. These externalities exist since the decision to invest in education increases individual productivity and hence wages, and may also have the additional effect of increas- ing the productivity of other workers. Despite this, there is no general scheme for financ- ing education. In some countries, investment in higher-level education is supported by

the state, while in others it is exclusively private. It is, therefore, highly relevant to verify whether educational subsidy, as dictated by policy, enhances the welfare of society.

Undoubtedly, investment in human capital creates great opportunity for people, families, firms and society as a whole. Such an investment is considered to be the sim- plest means to higher levels of social welfare.

Accumulation of human capital accelerates technological and economic growth. Edu- cation improves worker productivity and therefore has an influence on earnings.

However, the total gains from investment in education could exceed the human capital rate of return. These gains may then lead to many positive externalities for society, for example, better hygiene and health standards. Educated people are presumed to be innovative and those less educated often fol- low their new habits and lifestyle. Educated members of societies have better capacity for understanding and processing new information and transmit this information to others. Therefore, it is rather obvious that

The present research was co-funded by the European Com- mission under the 6^th Framework Programme’s Research Infrastructure Actions (Trans-national Access contract RITA 026040) hosted by IRISS-C/I at CEPS/INSTEAD, Differdange (Luxembourg). The views expressed in this article are those of author and shall be not related to in- stitutions. The first version of this article was published as the IRISS working paper (www.iriss.ceps.ln/documents/

irisswp107.pdf).

* Addess: ul. Długa 44/50, 00-241 Warszawa, Poland.

E-mail: pstrawinski@wne.uw.edu.pl.

(2)

118 ^Strawiński external effects of education must exist;

but, these external effects are challenging to quantify.

Many studies have shown that an additional year of schooling increases an individual wage by 5–10%; even up to 15% in countries with a relatively low GDP per capita during economic expansion (Psacharopou- los and Patrinos, 2002). However, the economic consequences of change in the mean number of years in education may be different to this human capital return. A positive change in average level of education increases the skilled-work supply, and could also influence labour demand. Such growth may increase total wages and human capital returns to education for two separate reasons. First, the standard neoclassical model suggests that, if educated and uneducated workers are imperfect substitutes, an increase in the proportion of educated workers will raise wages for both groups. Secondly, a human capital spillover may result from the exchange of ideas and learning by doing. Those with a lower level of education may acquire skills simply by imitation of highly educated workers. The increase in wages is an effect of increased productivity; however, if education also has a signalling effect, or if supply of other production factors is inelastic, this increase in wages is lower than the human capital rate of return on education. There- fore, the value of education to a society may exceed the rate of human capital return as a result of the positive external returns. De- spite its potential significance to economic policy, much less is known about the external return to education than to human capital.

The concept of external return to education was brought into economic analysis by Acemoglu and Angrist (1999), and their approach was extended by Moretti (2004).

The analysis departs from the concept of social capital. Bourdieu (1986) raised the argu- ment that “social capital is an attribute of an individual in a social context. One acquires social capital through purposeful action and

can transform social capital into convention- al economic gain”. In this context, social return to education may be defined as the part of the return that can be attributed to social capital. Social effects increase the return on education, but cannot be captured in a standard human capital-based framework.

In order to formalise this concept, Moretti (2004) defined social return to education as the sum of human capital return and external return. The former is often treated as a private return. The latter is defined as the effect of an increase in the proportion of educated workers in the area on the total sum of wages less the change in human capital return. For the reasons already mentioned, private and external returns should be estimated simultaneously. However, there is no straightforward direct measure that captures the external return to education. The usual proxy suggested in the literature is the effect on wages caused by the increased numbers of educated workers.

Many economists have studied human capital returns to education. Several economic surveys have found a positive relationship between an educational degree and salary received. Labour market research indicates that in the United States, each additional year of education increases an average wage by 7.5% (Acemoglu and Angrist, 1999). In a similar article, Blundell et al.

(2005) showed, using various econometric techniques, that a degree raises the average salary by 25% in the United Kingdom.

Similar results have been obtained in studies of other European countries, where the estimated rate of return ranges between 7 and 12% (Psacharopoulos and Patrinos, 2002).

Brunello et al. (2001) examined Italian labour market data and showed that the average yearly rate of return on university education was approximately 6.2% for males and 7.5% for females. Comparable results for the EU-15 were obtained by Harmon et al.

(2002). They estimated the average annual rate of return at 6.5%. De la Fuente (2003),

(3)

External return to education in europe 119 in a report prepared for the European Com-

mission, estimated the yearly rate of return to education at 6.2% for the “old” EU members, while he stressed that in the long term there is an additional 3.1% premium from quicker technological development.

The empirical literature on social returns to education is rather limited; evidence is provided only by a few studies. Studies at a microeconomic level report individual log wages by individual years of education, average years of education in the relevant geographical area of interest and additional control variables. The social returns are the sum of the two education coefficients: one for human capital return and the other for external return. Rauch (1993), in one of the few studies on the topic, found evidence for an 8.1%

social rate of return with a 3.3% external rate in the United States by comparing wage increases with average educational attainment in an area. Acemoglu and Angrist (1999) estimated social return as approximately 7.5%

(external 4.6%) using ordinary least squares (OLS) and the instrumental variable method, 9.1% with a 1.8% external rate. Moretti (2004) estimated spillover from college education by comparing wages for otherwise similar individuals working in cities with contrasting proportions of college graduates in the labour force. They found a positive, significant relationship between an increased supply of college graduates and average wages. How- ever, all these studies are limited to the U.S.

labour market.

The other branch of this research involves sector analysis within an industry. Sekelle- toriu and Maysami (2004) studied this type of external effect in Latin American countries and found a positive external effect of 2–4%. In the United Kingdom, Kirby and Riley (2008) found a positive external return to education at industry level. The margin is approximately 3%, which is comparable to previous findings.

The macroeconomic approach to return on education uses cross-country regression

and uses the log of GDP per capita explained by average schooling and additional control variables. Heckman and Klenow (1997) com- pared the schooling coefficient from the human capital model with one calculated according to a macroeconomic model; their estimate of social return was 10.6%. Bils and Klenow (1998) used a similar approach.

When they accounted for differences in technology, the magnitude of social returns ap- proached private returns and external return disappeared. In a similar study, Topel (1999) also used cross-country regressions and estimated an external return to education of 6.2%.

In this study, the following question is asked: “Is the level of external return similar in all European countries?” If the external return were to vary, it would prompt the further question: “Does this variation equalise the social return to education between countries or does it contribute further to differences in profitability yielded by higher education in different countries?

This has significance for policy-making.

If the social return on education is similar across Europe, there should be no direct economic incentives to emigrate to obtain higher returns, and consequently better living conditions. If the opposite is the case, after labour market liberalisa- tion, one might expect a greater tendency towards migration among well-educated.

The model is based on the comparative advantage theory, and Mincerian wage equations are estimated. In order to as- sure robustness of the main results, two different datasets are explored: the Con- sortium for Household Panels for Socio- Economic Research (CHER) for 1990–2000 and the European Community Household Panel (ECHP) for 1994–2001. The novelty of this study is that internationally harmonised data is used, with additional effort to control for the potential endogeneity of decisions about education, average schooling and self-selection.

(4)

120 ^Strawiński

Methodology

Measuring social capital is an ambiguous task. There is no widely held consensus on how to measure it, or at which level, macro or micro. The between-country regression approach usually neglects differences in technology level, or uses coarse proxies.

Moreover, within-country exceeds between- -country variation for education. For this reason, macro-level estimates tend to be sig- nificantly higher than those at an individual level. Psacharopoulos and Patrinos (2002) warned that overall results using the macroeconomic approach are inconclusive.

The underlying problem with the microeconomic approach is that the factors po- tentially responsible for creation of external return to human capital are not easy to quantify. Some methods proposed in the literature suggest that level of education of the popu- lation and its geographical concentration are good proxies for associative behaviour;

therefore, they can be measures for the external effect as a part of social capital (Moretti, 2004).

There are several ways of estimating the rate of return to education. In this study, the Mincer human capital model was employed (1974). This is the most fre- quently used model in empirical economics.

The Mincerian wage equations are commonly used in several areas of labour economics, such as return to education, wage inequali- ties, or the pay–gender discrimination gap.

This method entails fitting empirical data to the logarithm of the actual wage by linear regression. Characteristics such as level of education, age as a measure of work experience and socio-demographic characteristics are used as explanatory variables.

The analysis of the social return on education, in addition to a human capital rate of return to education, must accommodate educational spillover effects. Education may affect national income in ways that are not fully measurable by wages. For example,

education is positively correlated with level of participation in the labour force. Several aspects of daily life, including health and safety standards, electoral participation, and voting behaviour are influenced by a society’s educational level. For example, in de- veloping countries, education is negatively associated with women’s fertility and positively associated with infant health (Kreuger and Lindahl, 2001). The more educated societies are, the better they understand the interde- pendencies between different aspects of life, and the better the collective decisions they take. These indirect effects are a vital part of social return. Moretti (2004) formulated a theoretical framework that allows for social return. In his general equilibrium model, an increase in the number of educated workers in the local labour market may raise the aver- age wage above the private return on schooling, even in the absence of any spillover. This is the case in a market with a high intensity of highly skilled workers. The concern is that, according to this model, individuals in regions with high human capital are in- herently better workers than individuals with the same observable characteristics who live in regions with low human capital intensity.

This situation leads to a self-selection problem, as predicted by the Roy model of self- -selection. According to this model, people with similar social and demographic back- grounds are more likely to take up education if they live in a region with a high intensity of skilled labour.

Our empirical approach is similar to that of Acemoglu and Angrist (1999). We define social return to education as the sum of human capital return to education and the indirect effect of an increase in the proportion of educated workers on wages. The latter is called the external wage effect in the literature. This effect is equal to the effect of an increase in the proportion of educated workers minus the effect of private returns to education. The model itself is based on the comparative advantage theory. Individuals choose

(5)

External return to education in europe 121 their preferred education level. In order to do

that, they compare streams of future income with alternative education levels. They could withdraw from the education system at any moment. Continuing education is considered an investment because it entails choice between current costs and future income.

Education postpones entry into the labour market and reduces working activity time.

Analogously to the standard cost-benefit analysis of investment project, it is possible to calculate the internal rate of return.

In order to reduce the complexity of the analysis, the rate of return to education is treated as the parameter characteristic of an individual. It is assumed that investment at an individual level has no impact on the general equilibrium of the economy. Thus, the mar- ginal return rate is not affected by the decision of other society members. The next simplifying assumption is that study costs are uniformly distributed over a study period.

In reality, they are usually higher at the be- ginning and then decline.

Let I_ij be the lifetime labour income of person i with education level j. Let X_i be a vector of observable abilities and socio- demographic characteristics and ε_i a vector of unobservable terms that have an influence on labour income. Then, the lifetime income is defined by:

(1) Assume that the cost of achieving educa- tion level j for an individual i is equal to C_ij. It varies among individuals due to specific abil- ities and predisposition heterogeneity. Let V_ij be a value of utility function derived for per- son i from an education level j. The mecha- nism of choosing the desired education level can be represented as:

(2) It is presumed that people behave according to a maximum utility theory. Therefore,

one chooses education level j that maxim- ises the difference between future incomes attached to this level and the cost required to achieve it.

The analytic formula is an extension of Willis and Rosen’s (1979) model of demand for education combined with Moretti’s (2004) approach. From the former, we borrow the selection mechanism, and from the latter, an additional regressor for education in the local area. In our model, in addition to human capital return, we also consider social return to education. We distinguish between highly skilled workers H and lower-skilled ones L. We emphasise return to the secondary (high school or equivalent) and tertiary levels (university or equivalent) of education. At the first stage of education, primary school is compulsory;

therefore, the lack of an appropriate comparison group makes a return calculation for that education level impossible. Each education level has its own initial earn- ings: w_L0for secondary education and w_H0 for tertiary. We assume that wages are in- creasing functions of time. The rate of wage growth g depends on skills achieved dur- ing the education process and equals g_H for a person with a higher skill level (university or equivalent education in case of return on tertiary education, or high school or equiva- lent in case of secondary education) and g_L for workers with fewer skills. The schooling process is time-consuming. In order to at- tain a higher degree of education, a person has to sacrifice some of his potential labour activity time. The amount of time necessary to achieve a degree is represented as T years.

If one chooses a higher level of education, his future stream of incomes w_Hi is given by

(3)

The variable t represents working time and (t – T) is a measure of work experience.

42 ^Strawiński

their preferred education level. in order to do that, they compare streams of future income with alternative education levels. They could withdraw from the education system at any moment. Continuing education is considered an investment because it entails choice between current costs and future income.

in order to reduce the complexity of the analysis, the rate of return to education is treated as the parameter characteristic of an individual. it is assumed that investment at an individual level has no impact on the general equilibrium of the economy. Thus, the mar- ginal return rate is not affected by the decision of other society members. The next simplifying assumption is that study costs are uniformly distributed over a study period.

in reality, they are usually higher at the be- ginning and then decline.

Let I_ij be the lifetime labour income of person i with education level j. Let X_i be a vector of observable abilities and socio- demographic characteristics and ε_i a vector of unobservable terms that have an influence on labour income. Then, the lifetime income is defined by

(1) Assume that the cost of achieving educa- tion level j for an individual i is equal to C_ij. it varies among individuals due to specific abil- ities and predisposition heterogeneity. Let V_ij be a value of utility function derived for per- son i from an education level j. The mecha- nism of choosing the desired education level can be represented as

(2) it is presumed that people behave according to a maximum utility theory. Therefore,

The analytic formula is an extension of Willis and Rosen’s (1979) model of demand for education combined with Moret- ti’s (2004) approach. From the former, we borrow the selection mechanism and from the latter, an additional regressor for education in the local area. in our model, in addition to human capital return, we also consider social return to education. We distinguish between highly skilled workers H and lower-skilled ones L. We emphasise return to the secondary (high school or equivalent) and tertiary levels (university or equivalent) of education. At the first stage of education, primary school is compulsory;

therefore, the lack of an appropriate comparison group makes a return calculation for that education level impossible. Each education level has its own initial earn- ings: w_L0for secondary education and w_H0 for tertiary. We assume that wages are in- creasing functions of time. The rate of wage growth g depends on skills achieved dur- ing the education process and equals g_H for a person with a higher skill level (university or equivalent education in case of return on tertiary education, or high school or equiva- lent in case of secondary education) and g_L for workers with fewer skills. The schooling process is time-consuming. in order to at- tain a higher degree of education, a person has to sacrifice some of his potential labour activity time. The amount of time necessary to achieve a degree is represented as T years.

if one chooses a higher level of education, his future stream of incomes w_Hi is given by

(3)

I_ij = f(X_i,ε_i).

V_ij = max(I_ij– C_ij).

j

w_Hi(t) = �w_Hoexp(g0_H(t – T)) T < t < ∞0 ≤ t ≤ T .

42 ^Strawiński

(3)

I_ij = f(X_i,ε_i).

j

42 ^Strawiński

(3)

I_ij = f(X_i,ε_i).

j

(6)

122 ^Strawiński We can describe an income equation for a low-educated person in a similar manner:

(4) An income stream is determined by two parameters: the starting salary for each education level w_.0 and the growth rate g.

A person making a decision regarding a desired education level compares discounted future values of potential income. The per- son i chooses university education if the net benefits from achieving a higher degree are greater than the benefits from a lower level of education.

The discounted values of an education level reflect the economic mechanism of choosing between two different education levels. The salary level is a function of education, experience measured by age and social and demographic characteristics. In labour economics, it is commonly assumed that the distribution of earnings is well ap- proximated by the log normal distribution.

The wage equation for each education level could be represented by the classical linear regression model. Following Acemoglu and Angrist (1999) and Moretti (2004), we allow for human capital spillover by letting worker productivity depend on the proportion of educated workers in the local labour market.

a human capital quality measure is added to the following wage equation:

(5)

where X_i is a matrix of socio-demographic characteristics including working experience and its square, betas are wage equa- tion coefficients, Y_edui is the number of years spent in an education system (education level proxy), H is the number of highly-skilled workers in a local labour market, and L is the size of the local lesser-skilled labour force.

The coefficient is an estimate of the average

yearly return on schooling and is the proxy for the external effect. The latter is the coefficient of interest, which is the estimate of the effect of the proportion of those with higher education on average wages after controlling for private returns to education.

As indicated by Moretti (2004), the wages of uneducated workers benefit from an increase in the proportion of educated workers for at least two reasons. First, an increase in the number of educated workers raises uneducated worker productivity because of imperfect substitution. Second, the spillover raises their productivity further.

The principal challenge in estimating a causal effect of education on wages is iden- tification. Individual education and average schooling levels could both be correlated with wages for various reasons; thus, the observed relationship between variables is not necessarily causal (Acemoglu and Angrist, 1999). The education level, up to a point, is pre-determined by the social background of the person (Becker, 1976). It is more likely that a person’s decision to study is positively affected by living in an area where most of the people are highly educated. As a result, an individual’s education and the average education are possibly correlated. Moreover, educational decisions depend on the ability of the person, something which is not directly observed. Therefore, the process of choosing a desired education level could be treated as self-selection. Moreover, as is shown in many studies, individual wages are related to an unobserved characteristic (i.e. ability). There could be an endogeneity problem and a potential sample selection problem. The schooling decision could be endogenous due to the fact that education of the individual influences the proportion of highly skilled people in the area; at the same time, those living with highly educated neighbours are more likely to obtain higher education that those who live in other communities. The reasons for potential self- -selection are twofold. The first involves External Return to Education in Europe 43

We can describe an income equation for a low-educated person in a similar manner:

a person making a decision regarding a desired education level compares discounted future values of potential income. The per- son i chooses university education if the net benefits from achieving a higher degree are greater than the benefits from a lower level of education.

The discounted values of an education level reflect the economic mechanism of choosing between two different education levels. The salary level is a function of education, experience measured by age and social and demographic characteristics. in labour economics, it is commonly assumed that the distribution of earnings is well ap- proximated by the log normal distribution.

a human capital quality measure is added to the following wage equation

(5)

where Xi is a matrix of socio-demographic characteristics including working experience and its square, betas are wage equa- tion coefficients, Y_edui is the number of years spent in an education system (education level proxy), H is the number of highly-skilled workers in a local labour market, and L is the size of the local lesser-skilled labour force.

As indicated by Moretti (2004), the wages of uneducated workers benefit from an increase in the proportion of educated workers for at least two reasons. First, an increase in the number of educated workers raises uneducated worker productivity because of imperfect substitution. second, the spillover raises their productivity further.

The principal challenge in estimating a causal effect of education on wages is iden- tification. individual education and average schooling levels could both be correlated with wages for various reasons; thus, the observed relationship between variables is not necessarily causal (Acemoglu and Angrist, 1999). The education level, up to a point, is pre-determined by the social background of the person (Becker, 1976). it is more likely that a person’s decision to study is positively affected by living in an area where most of the people are highly educated. As a result, an individual’s education and the average education are possibly correlated. Moreover, educational decisions depend on the ability of the person, something which is not directly observed. Therefore, the process of choosing a desired education level could be treated as self-selection. Moreover, as is shown in many studies, individual wages are related to an unobserved characteristic (i.e. ability). There could be an endogeneity problem and a potential sample selection problem. The schooling decision could be endogenous due to the fact that education of the individual influences the proportion of highly skilled people in the area; at the same time, those living with highly educated neighbours are more likely to obtain higher education that those who live in other communities. The reasons for potential self-selection are twofold. The first involves w_Li(t) = w_Loexp( g_Lt) 0 ≤ t < ∞ .

ln(w) = Y_{edu i}γ + Xʹ_iβ + �H + L � δ + ε .H

External Return to Education in Europe 43 We can describe an income equation for

a low-educated person in a similar manner:

a person making a decision regarding a desired education level compares discounted future values of potential income. The per- son i chooses university education if the net benefits from achieving a higher degree are greater than the benefits from a lower level of education.

The discounted values of an education level reflect the economic mechanism of choosing between two different education levels. The salary level is a function of education, experience measured by age and social and demographic characteristics. in labour economics, it is commonly assumed that the distribution of earnings is well ap- proximated by the log normal distribution.

a human capital quality measure is added to the following wage equation

(5)

where Xi is a matrix of socio-demographic characteristics including working experience and its square, betas are wage equa- tion coefficients, Y_edui is the number of years spent in an education system (education level proxy), H is the number of highly-skilled workers in a local labour market, and L is the size of the local lesser-skilled labour force.

As indicated by Moretti (2004), the wages of uneducated workers benefit from an increase in the proportion of educated workers for at least two reasons. First, an increase in the number of educated workers raises uneducated worker productivity because of imperfect substitution. second, the spillover raises their productivity further.

The principal challenge in estimating a causal effect of education on wages is iden- tification. individual education and average schooling levels could both be correlated with wages for various reasons; thus, the observed relationship between variables is not necessarily causal (Acemoglu and Angrist, 1999). The education level, up to a point, is pre-determined by the social background of the person (Becker, 1976). it is more likely that a person’s decision to study is positively affected by living in an area where most of the people are highly educated. As a result, an individual’s education and the average education are possibly correlated. Moreover, educational decisions depend on the ability of the person, something which is not directly observed. Therefore, the process of choosing a desired education level could be treated as self-selection. Moreover, as is shown in many studies, individual wages are related to an unobserved characteristic (i.e. ability). There could be an endogeneity problem and a potential sample selection problem. The schooling decision could be endogenous due to the fact that education of the individual influences the proportion of highly skilled people in the area; at the same time, those living with highly educated neighbours are more likely to obtain higher education that those who live in other communities. The reasons for potential self-selection are twofold. The first involves w_Li(t) = w_Loexp( g_Lt) 0 ≤ t < ∞ .

ln(w) = Y_{edu i}γ + Xʹ_iβ + �H + L � δ + ε .H

(7)

External return to education in europe 123 the influence of a person’s ability on their

schooling decision. Unfortunately, ability is not directly observed. Otherwise, we find no support in the literature for the sugges- tion that people in one particular area could have more educational talent than popula- tions elsewhere. The second reason involves the individual’s behaviour in the labour market. In order to calculate the returns, one needs to observe wages. However, the opportunity cost of work differs between countries.

In the presence of an endogeneity or self- selection problem, standard estimators would be inconsistent (Blundell et al., 2005). In order to alleviate the endogeneity problem, an instrumental variable approach must be used;

that is, education in the area has to be replaced by set of instruments, and in the case of selection bias, it is necessary to include a selection equation in the model. This equation describes the mechanism for selecting the observations–working people–for the estima- tion sample. The non-labour income is used to identify selection equation and regional dummies are used only in selection equation.

The complete model can be expressed as

(6)

where w₀ is the selection indicator interpreted as a wage offer above a reservation wage, Z_i is the selection variable matrix, and alphas are selection equation coefficients.

Data

Two broad datasets were used to analyse the external effect of education. The first, the ECHP, was a harmonised European longitudinal survey of households, income, and living conditions. The survey ran from 1994 to 2001 and encompassed 15 EU member states (the EU-15 or “old member states”). The sample for each year comprised

information on approximately 65 000 households and 130 000 adults (170 000 individuals including children). Unfortunately, it was not possible to take advantage of the longitudinal nature of the data in the analysis for two reasons. Firstly, for most countries, the data had a rolling panel design; secondly, the results for the remaining countries might have been heavily influenced by attrition bias¹.

The second source of empirical data was CHER, an internationally comparative microeconomic database that integrated longitudinal datasets from Europe and the United States over a many years (1990–2001) and included countries in the ECHP. However, for most countries, data were available from 1994. The base contains data for 18 countries (14 EU members in 1994, as well as Switzer- land, Poland, Hungary, and the United States). Topics encompassed by the data are labour force participation and related issues, income components, and social relation- ships. These data contain the averages of approximately 75 000 households and 150 000 individuals surveyed annually; however, the number of participating countries varied from year to year².

Both datasets mentioned are from national surveys; therefore, not all relevant variables for the analysis were available for all countries. This issue is discussed later.

Here, it was decided to use both datasets for several reasons. The ECHP data contained more observations for each country, while the CHER data covered a larger number of countries. Pursuing the analysis on both datasets provided a simple robustness check.

The list of countries and sample sizes is presented in Table 1.

The empirical sample is restricted in terms of several dimensions. First, analogously to Kirby and Riley (2008), the analysis was narrowed to individuals aged 30 to 55. Younger

1 The attrition rate is approximately 15%.

2 We excluded the United States from the CHER sample as it is a non-European country.

44 ^Strawiński

the influence of a person’s ability on their schooling decision. Unfortunately, ability is not directly observed. Otherwise, we find no support in the literature for the sugges- tion that people in one particular area could have more educational talent than popula- tions elsewhere. The second reason involves the individual’s behaviour in the labour market. in order to calculate the returns, one needs to observe wages. However, the opportunity cost of work differs between countries.

in the presence of an endogeneity or self-selection problem, standard estimators would be inconsistent (Blundell et al., 2005).

in order to alleviate the endogeneity problem, an instrumental variable approach must be used; that is, education in the area has to be replaced by set of instruments and in the case of selection bias, it is necessary to include a selection equation in the model. This equation describes the mechanism for selecting the observations–working people–for the estima- tion sample. The non-labour income is used to identify selection equation and regional dummies are used only in selection equation.

The complete model can be expressed as

(6)

where w₀ is the selection indicator interpreted as a wage offer above a reservation wage, Zi is the selection variable matrix, and alphas are selection equation coefficients.

Data

Two broad datasets were used to analyse the external effect of education. The first, the ECHP, was a harmonised European longitudinal survey of households, income, and living conditions. The survey ran from 1994 to 2001 and encompassed 15 EU member states (the EU-15 or “old member states”). The sample for each year comprised

information on approximately 65 000 households and 130 000 adults (170 000 individuals including children). Unfortunately, it was not possible to take advantage of the longitudinal nature of the data in the analysis for two reasons. Firstly, for most countries, the data had a rolling panel design; secondly, the results for the remaining countries might have been heavily influenced by attrition bias¹.

The second source of empirical data was CHER, an internationally comparative microeconomic database that integrated longitudinal datasets from Europe and the United states over a many years (1990–2001) and included countries in the ECHP. However, for most countries, data was available from 1994. The base contains data for 18 countries (14 EU members in 1994, as well as swit- zerland, Poland, Hungary, and the United states). Topics encompassed by the data are labour force participation and related issues, income components, and social relation- ships. These data contain the averages of approximately 75 000 households and 150 000 individuals surveyed annually; however, the number of participating countries varied from year to year².

Both datasets mentioned are from national surveys; therefore, not all relevant variables for the analysis were available for all countries. This issue is discussed later.

Here, it was decided to use both datasets for several reasons. The ECHP data contained more observations for each country, while the CHER data covered a larger number of countries. Pursuing the analysis on both datasets provided a simple robustness check.

The list of countries and sample sizes is presented in Table 1.

The empirical sample is restricted in terms of several dimensions. First, analogously to Kirby and Riley (2008), the analysis was narrowed to individuals aged 30 to

1 The attrition rate is approximately 15%.

2 We excluded the United states from the CHER sample as it is a non-European country.

�ln(w) = Y_{edu i}γ + X_iβ + �H + L � δ + εH .

w₀ = Zʹ_iα + ξ_i

(8)

124 ^Strawiński

people were omitted in order to avoid bias from the direct influence of their schooling decisions on earnings. In some countries, the first job contract is limited by law.

In the case of older workers, an attempt was made to exclude the influence of retirement decision. Second, people who received income from work or self-employment were investigated. In addition, information was rejected on part-time employees, people with combined incomes from employment and social assistance or those whose work was not their main source of income. This step was necessary because the data did not provide exact hours worked; so, it was not possible to calculate hypothetical full-time earnings.

In addition, all groups of workers mentioned may have decided to work on a non- -earnings basis; their wages may not have reflected the actual value of their work abilities. In the self-selection specification, sources of income were controlled directly through a selection equation. The next restriction involved farming income, highly correlated

with land productivity and very weakly related to human capital productivity. As a con- sequence, a farmer’s income could only be partly determined by education and abilities.

In order to overcome the problem of eventual bias, the data was rejected from households for which farming was the only or the main income source. Handling the problem in this manner is justified in economic theory.

The additional restrictions originated from the data availability issue. Missing information on labour income is controlled by selection. In order to construct a proxy for measuring spillover effects, the data sets on education level and location of the residence (NUTS³ and town size variables) require ex- ploration. Unfortunately, some data were not available for a significant number of countries in both samples. NUTS information, when present, was only available at the NUTS1 level, which encompassed a large area of 3–7

3 NUTS: Nomenclature of Units for Territorial Statistics, which are territory units used by Eurostat.

Table 1 Sample sizes

Sample ECHP CHER

full NUTS NUTS&town Full NUTS NUTS&town

Austria 44 909 15 007 11 945 45 920 17 412 4 231

Belgium 40 698 16 815 13 267 48 344 13 319 13 304

Finland 41 831 18 854 13 725 41 982 20 706 0

France 86 770 32 818 26 243 95 171 33 933 7 242

Greece 83 276 25 073 19 567 85 748 30 356 6 738

Hungary – – – 25 668 8 088 4 604

Ireland 44 171 13 035 7 768 53 116 16 982 2 489

Italy 122 429 43 032 33 794 129 151 46 094 0

Portugal 87 682 29 524 25 335 91 437 34 935 8 451

Poland – – – 41 776 10 364 7 552

Spain 114 566 33 766 24 579 115 779 36 856 0

Sweden 45 177 22 863 20 536 – – –

United Kingdom 67 790 28 567 23 622 103 498 36 408 31 319

Based on ECHP and CHER data.

(9)

External return to education in europe 125

million inhabitants. For these reasons, small countries, like Luxembourg and Denmark, comprised only one NUTS. In addition, for Finland and Sweden, the NUTS variable was not available in the CHER sample. For several countries: Germany, Italy, the Netherlands, Spain, and Sweden in both samples the town size variable was not available for reasons of data anonymity. More- over, the variable definition differs between the ECHP and CHER data. In the former, the town variable may take three distinct

values, while in the latter only the urban–

–rural indicator is available. In addition, in- complete observations were discarded.

Being aware of major sample reductions, particularly in terms of the number of countries examined, it was decided to adjust the empirical strategy to available data and perform the analysis at two levels of data disaggregation: NUTS and town, and NUTS level. In the former, educational profiles were calculated for the areas defined by the prod- uct of NUTS and the town variable, while Table 2

ECHP sample average characteristics for 1998

Country Log

wage Gender Work

experience Years

of education Family Public

employment Self employment

Austria 9.73 0.35 21.22 12.08 0.62 0.27 0.06

0.63 0.48 9.49 1.94 0.49 0.44 0.24

Belgium 9.73 0.36 20.02 13.57 0.68 0.23 0.06

0.65 0.48 10.00 3.25 0.46 0.42 0.24

Finland 9.73 0.47 22.43 13.38 0.63 0.37 0.06

0.79 0.50 10.04 3.12 0.48 0.48 0.24

France 9.74 0.40 22.25 12.34 0.66 0.31 0.07

0.66 0.49 10.05 3.24 0.47 0.46 0.26

Greece 9.31 0.34 21.63 12.46 0.75 0.28 0.26

0.66 0.47 10.40 3.23 0.44 0.45 0.44

Ireland 9.58 0.32 20.55 12.34 0.68 0.33 0.12

0.66 0.47 9.96 2.99 0.46 0.47 0.32

Italy 9.46 0.35 22.99 11.26 0.71 0.29 0.20

0.74 0.48 9.79 2.54 0.45 0.45 0.40

Portugal 9.09 0.41 22.89 10.48 0.75 0.20 0.13

0.79 0.49 10.07 2.76 0.43 0.40 0.33

Spain 9.48 0.32 21.42 12.14 0.68 0.22 0.14

0.78 0.47 10.55 3.49 0.47 0.41 0.35

Sweden 9.29 0.45 22.89 13.30 0.25 0.21 0.03

0.73 0.50 10.60 2.92 0.44 0.40 0.18

United Kingdom 9.65 0.37 21.38 13.40 0.65 0.24 0.12

0.72 0.48 10.76 3.69 0.48 0.43 0.32

Based on ECHP data. The numbers in top row for each country represent mean value of characteristics in the sample;

the numbers appearing on the bottom line are standard deviations.

(10)

126 ^Strawiński

in the latter they were defined by NUTS only. It was known that the latter method of data preparation might be not suitable to capture the external effect, if present. In this approach, it was assumed that within one NUTS, concentration of human capital only depended on size of town.

The datasets did not directly provide information on years of schooling. For the pur- pose of the analysis, this was obtained by imputation using information contained in the achieved education level. In order to obtain results that were comparable for different countries, the assumptions were made that ISCED level 2 required 9 years

of education, level 3 required 12 years and levels 5 or 6 require 17 years of education.

Before the return rate of university was calculated, the basic sample characteristics for each country were analysed. This data are presented for the NUTS level, because when a town variable was collected, it was very rarely missing. After all data correc- tion operations, approximately a third of the initial observations remained in the sample for each country. The main characteristics for the EHCP sample are presented in Table 2 and those for the CHER sample are presented in Table 3. The values are for the year 1998, but are fairly similar for the other Table 3

CHER sample average characteristics for 1998 Country Log wage Gender Work

experience Years

of education Family Public

employment Self employment

Austria 9.50 0.35 21.54 11.99 0.89 0.25 0.14

0.54 0.48 9.47 1.88 0.32 0.43 0.35

Belgium 9.62 0.45 20.62 13.19 0.89 0.23 0.15

0.74 0.50 9.69 3.31 0.32 0.42 0.35

France 9.12 0.56 22.97 12.34 0.88 0.22 0.06

0.92 0.50 12.97 3.65 0.33 0.41 0.24

Greece 9.24 0.49 22.52 12.13 0.93 0.24 0.32

0.63 0.50 10.79 3.13 0.26 0.43 0.47

Hungary 7.95 0.48 22.12 11.46 0.96 0.25 0.06

0.66 0.50 8.57 2.92 0.20 0.43 0.24

Ireland 9.77 0.31 21.20 12.11 0.91 0.29 0.15

0.62 0.46 10.16 2.98 0.29 0.45 0.36

Italy 9.51 0.33 23.11 11.16 0.93 0.28 0.25

0.54 0.47 9.69 2.49 0.25 0.45 0.43

Poland 7.40 0.48 24.33 10.87 0.93 0.38 0.19

0.57 0.51 9.27 2.47 0.25 0.36 0.39

Portugal 8.90 0.43 23.26 10.48 0.98 0.17 0.18

0.74 0.49 10.45 2.81 0.15 0.37 0.38

Spain 9.35 0.38 22.05 12.03 0.96 0.19 0.21

0.82 0.49 10.47 3.44 0.20 0.39 0.41

Based on CHER data. The numbers in top row for each country represents mean value of the characteristics in the sample; the numbers in the bottom line are standard deviations. All numbers are for 1998, except Hungary (1997).

(11)

External return to education in europe 127

years. The year 1998 was chosen to maximise the number of countries in CHER sample.

The basic characteristics of both datasets are fairly similar. The dependent variable is the logarithm of yearly gross wages and salaries expressed in 2005 Euros. Yearly wages, rather than hourly rates, were used to eliminate the differences in working hours between countries. Original numbers in national currencies were converted to euros using the annual exchange rate published by Eurostat and deflated by the HCPI to be comparable between different countries.

For six countries, the averages of log wage were higher in the CHER sample and for two countries in ECHP sample. The differences arise from the random nature of rep- resentative samples. Fortunately, these differences did not influence the results as it was the relative percentage difference between workers with high and low education levels that were of interest, rather than their actual wages.

The sample averages for gender, number of years spent in education and years of work experience were fairly similar in both datasets. Noticeable differences were observed for family and self-employment indicators.

The difference in the family variable origi- nates from its construction. In the ECHP sample it is derived from the marital status variable, while in the case of the CHER sample it is derived from household size.

This was done purposely in order to reduce the problem caused by missing values for marital status in the CHER sample. The var- ying values for the self-employment dummy reflect a different definition of the variable.

Generally, the definition in the CHER sample comes from the labour activity section of the data and includes self-employed and self-employed with co-workers, whereas in the ECHP the self-employed are identified by declared main source of income.

Table 4 presents the educational structure in the analysed European countries.

Table 4

Educational structure in European countries (in %) Education

level Country

ECHP CHER OECD

Tertiary Secondary Tertiary Secondary Tertiary Secondary

Austria 6.0 59.9 6.0 59.8 11 56

Belgium 27.6 32.3 24.4 29.4 25 31

Finland 25.5 38.2 25.5 38.2 32 39

France 10.9 39.3 11.0 39.3 11 40

Greece 14.9 28.1 15.1 28.1 16 26

Hungary – – 10.3 52.2 13 50

Ireland 17.9 34.3 18.2 34.3 21 30

Italy 6.8 32.5 6.9 32.6 9 31

Poland – – 10.0 65.8 11 64

Portugal 7.0 12.4 7.0 12.3 9 11

Spain 26.4 58.3 26.7 58.1 24 57

Sweden 25.3 44.9 – – – –

United Kingdom 25.9 55.4 23.7 59.4 24 57

Source: Own computation based on ECHP and CHER data and OECD (2000).